At least with Erlang the edge cases won't take down the whole system... Erlangs paradigm was always 'Let it crash', in fact the whole language was designed with Fault tolerance in mind, they knew in critical systems like telecommunications the system needs to always remain up and available, so rather than attempt to defensively program, they viewed there will always be unforeseen bugs. hardware may fail, but they asked how can we ensure those problems never break the wider system. Every user action is isolated in its own PID, so when an edge case arises in that PID it just crashes, and a supervising process just restarts it.
They found all the mythical 10x engineers, captured them, and locked them in a basement with only laptops with Linux terminals and a mountain of anime pron to keep them company Those engineers have never been more happy
For context about FreeBSD, because the article skims over it: For years Erlang developers preferred FreeBSD because IO multiplexing is done in FreeBSD with something called kqueue, which requires one less syscall than the Linux equivalent (epoll). I believe Linux has iouring now. Let us just appreciate for a moment that the Erlang VM handles IO so well that the OPERATING SYSTEM ITSELF BECOMES THE BOTTLENECK.
A manager went to the Master Programmer and showed him the requirements document for a new application. The manager asked the Master: "How long will it take to design this system if I assign five programmers to it?" "It will take one year," said the Master promptly. "But we need this system immediately or even sooner! How long will it take if I assign ten programmers to it?" The Master Programmer frowned. "In that case, it will take two years." "And what if I assign a hundred programmers to it?" The Master Programmer shrugged. "Then the design will never be completed," he said.
Yea it’s not like this one big villain bro These are different people Then this L mentality when you’re seeking revenge because you were “rejected”…. Duh I’m pretty sure such intelligent and successful people are above that
Just a tiny technicality; Erlang uses "green" processes rather than green threads. Difference is that threads can share state, memory and address spaces, processes are isolated and share nothing (other than by copying values). Erlang implements its processes in its runtime system and is managed in its VM, rather than using OS threads or processes.
@@SeanPoulter Its amazing how easy is to create massively scalable things with Erlang, like most of the heavy lifting is really the programming paradigm. I was reading the Ejjaberd source code and I was impressed how little code in relation to the runtime is necessary to create the core of what is Whatsapp.
Try doing that with C you might save a lot of memory, but you lose all scaling as well. Its almost like C is cheating by being unsafe with its memory model full of aliasing and sharing via the use of pointers. You are also going to need 35.000 engineers because C is a shit programming language that requires a lot of stupid work to manage the memory. Its ironic that computer were invented to automate industrial process, yet, the very own engineers working with computers decided to not industrialize the most boring and stupid programming task : managing memory. All because in 1970, the PDP11 only have like 8Kb of RAM or something... But Erlang coming from the telephony world don't use memory to process data, only to control process that control the data flowing in actual wires, that's the magic. Its the control-plane/data-plane separation. Stored program with mutable shared memory and big linear address space was a mistake, and its not even how hardware works anymore anyways, the so called "C" programmers don't even program bare-metal anymore, its all a fad, its only there for their ego or something, that a modern compiler has to make a modern computer look like a 1970's era PDP11. And we are forever going to be stuck in this memory model thanks to C and Unix and Intel.
Agreed. I learned nothing from this video. Glad I skipped 90% of it. Dev youtubers really should focus more on providing value with their article-based videos. Most of the time it's empty phrases like "if you've never worked on x problem....it's HARD! Like...actually hard" and other trivial statements. Feel like it's more about engaging the dev community by addressing obvious common concerns, struggles and frustrating experiences as a dev more than providing new insights
This respond from you is that you have not being born before Internet was a thing, I worked with Video Editing/Rendering to built those system in the 1994, the hard drives needed a special software because normal hard drives where "calibrating" and in those moments you where losing frames when you rendering hard core stuff, now days, just put a SSD or NVMe drive and you think problem solve. That is what programming is today, put just more Framework, and in the end you have just built a system even with the newest hard drives/SAN, and your performance will be like if i remember 100 IOPS per hard drive, so in a stripe configuration for performance, you should know that would be a cascading effect, . But you got to love that bill from any of the Cloud providers. If you ever built a successful app. LOL and BTW this was the time when Silicon Graphics had there computers running the HIPPI protocol. To much knowledge have been lost in time. But one thing remains, we that have struggle during the early days, really know how to build system from ground up today.
Having worked at Ericsson, I can tell you that they mostly dont use Erlang these days. Where I worked there was a lot of Scala and Akka, which is funny to me. One really cool feature of Erlang is that they could predict the amount of hardware needed. X amount of messages means Y hardware requirements, 2X messages 2Y hardware. But the coolest feature has to be the suicide pacts of processes
Akka is cool, if only wasn't running under JVM, and if I'm not wrong is just replicating what Erlang already does. Can't figure out why would they do that.
I worked in an IoT platform, and we reached a peak that we were struggling to manage. We started talking to some people who managed a lot of messages a day, and they were shocked by our volume at the time I worked there. We were hitting 500 million a day, and passing that mark was a challenge. I cannot imagine how it is to handle 50 billion.
I was expecting more from the article tbh. Yeah, it lists some things, but it's just a lists of concepts that by now I think it's just standard good practices. (Aside from the lang choice and fBSD OS choice). I was expecting the article to go a little deeper. It's nice to know that they located and fix bottlenecks, but I wanted to know, how they identified, what was the issue and how that got fixed.
Fear of telling too much and then having the information used against you later; fear of clones; industrial secret; or it just takes too much effort to ask the techs how shit is actually working.
They have life and people who care how about you ? You need to comment here to express yourself ? No one left to listen you ? Oh you guys work for groceries worth $600 per month ,$2000 rent and and flex of earning $4000 per month? Thinking elite indians manage all this for $100 and still saves for life. So live a life
@@riwajghimire I was stating a fact. WhatsApp is the most common messaging service in the subcontinent. Maybe you should drink a chai and chill, Riwaj.
I think perhaps another advantage that WhatsApp has (or had) was lack or server side storage? Generally messages were just sent and lived on client devices. When you went to a new device, if you didn't do a transfer yourself, you'd start fresh. I'm sure there was temporary storage but when you can cut out those huge database challenges you can save more engineering resources
I have a theory that pretty much any scale of software system can be produced if you stick just a handful of engineers in a couple of rooms. If it's a hard problem, give them £300k plus and hire the very best. It will still work out cheaper than 200 drones grinding away in their corporate cubicles. The effort of coordinating all those low skilled hacks and the quality control bureaucracy you'll need will cost far more than simply hiring good people in the first place.
@@bos9824 According to the stats, £300k is around the median for a top-tier engineer in the US and £250k in the EU. If you want the best of the best, you'd have to pay a bit more than the median.
okay I think I still am a javascript advocate but @ThePrimeTime is like the Red pill I needed because "You just pushed back 591 messages because you were off by a half millisecond"
While Erlang is functional it is perhaps the least important fact about the language and given the syntax and style of Erlang it is very procedural feeling. I find C/C++/Go/Java devs have little issue moving to it once getting used to replacing loops with maps, folds, and tail recursive functions.
Maps as in applying functions to iterable objects? I learned that Go call key/val pairs for maps after i learned that - So i'm wondering what the actual norm of maps actually is
@@GOTHICforLIFE1 map as in the high order function. Like fold. Check wikipedia for "map (higher-order function)". It is a function / behavior. Not a data structure.
@@GOTHICforLIFE1 Both are map, same word used for different things. But I would say more people would use map for applying a function to every element of a collection and call an associative array a dictionary.
@@disguysnReally? I like its relative simplicity. It might be inspired by a rather unpopular and odd language (Prolog) I really never minded it. I certainly prefer it over a number of other functional languages... including OCaml. I think the pattern matching syntax is the nicest I've seen in any language.
@7:40 No, not really green threads. That article misrepresents what is going on in Erlang. Erlang is using actor processes and has practically a full OS like process scheduler. And because it is built into the language and platform it offers features not found elsewhere and often can't be found elsewhere when the feature is an afterthought.
One important thing the author has missed: FreeBSD. it was very important for them to run FreeBSD for two reasons. 1. The co-founders were very familiar with the internals of FreeBSD. 2. The FreeBSD network stack was (and still is) way better than other open-source operating systems (e.g. Linux). They were able to support 2-3 million concurrent connections on a single commodity server.
Joe Armstrong is such an inspirational talker and person. In many ways whatsapp was the perfect showcase of the core ideas of what Erlang was designed for. Instant messaging needs to be low latency, highly reliable and capable of handling partial failures. And it was pretty well defined: Do SMS but over IP, with few new capabilities such as user-defined profile pictures.
This title sounds funny. Imagine 32 engineers sitting here and manually sending all those messages. And also it's sad that Telegram now goes on this path of "feature creep". It was a wonderful messenger and now it's full of social networking shit.
Thank you for showing where that meme came from, and the potential rapidity of meme generation and sharing. Love it! Am familiar with GIMP but your demo (about halfway through) produced much faster results.
@@ThePrimeTimeagen Netflix does contribute to the source code of FreeBSD, they got even more performance out of the network stack ! Thanks for your reply, I’m honoured 😊
Prime on CI/CD: "Everyone does that." (2008 btw) On the other end, here I am, trying to convince the company to let me implement pipelines in the repos while waiting for my boss to "deploy" with one-week delays using a copy-paste bash script, saying that it's good enough: 🤡
SMS in Ukraine is still expensive. We pay for every 100-200 messages, depending on provider. And 100 messages is equivalent in price to 50Gb of internet, on my provider.
@@catto-from-heavenFortunately I am not paying for them. But there are not that many plans, that do not have SMS included. There is an option to make your own plan, but its base cost is higher, than my whole plan. So, we just look for what is closest to what we need. And even then, we still get some things, that we do not use, nor need, but are still forced to pay for. My plan, for example, has 250 minutes per 4 weeks, to call other providers' numbers. Even if I use 10-12 of them at most, I still pay for them, as there is just no better plan, that suits my needs. And there are plans, that have 400 and even 600 SMS, that cost 8-12 times of my plan. There's a lot of them actually. If you have good internet - it's 99% that you'll get a ton of useless SMS with it. I literally never even once heard of anyone, who would use SMS, even if their plan provides them (and it usually does). Though, I get 3-4 daily spam ads via SMS. So yeah, people do not use it. The ad companies do. But we are still paying for them.
A part of how scalable the backend system is, we miss an important point here: the offline-first app. It's very hard to build an offline-first app even today, where we have the majority of problems already solved by third-party SDKs and packages. Building a seamless offline app in 2008 is some level of engineering to marvel at.
The "-agen" bit seriously never gets old. Also thank you for inspiring me to get back into coding a bit. Starting back with python and became enraptured by using python for AI stuff. Huggingface is awesome lol
I think that most shops should probably have between 3 and 40 devs depending on the depth of their app ecosystem. It doesn't make sense to have 600 or 40000 developers. That makes communication impossible...
3:33 - True, with all due respect, at the end of the day it is a messaging app with not a whole lot of features. By its very nature it is easy to parallelize and scale. Most messages are sent between 2 people and the group limit is 1024 (compared to other apps such as Telegram which support 200.000 or Discord which supports 250.000 members or more). The app existed for over 14 years now, so plenty of development time went into it already. 32 engineers (especially talented ones) is a lot for a problem as narrow as this. Past a certain point you simply can't split responsibilities anymore and you don't wanna end up having 3 senior engineers solely responsible for styling the Send button.
@@joaovmlsilva3509 I didn't assume there were 32 people just for the client, 32 engineers for the whole operation is a lot. And pretty sure they didn't require infrastructure to support 50 billion messages / day from launch day. Back in 2009 very few had smart phones world-wide. Now even people in starving 3rd world countries use smartphones and Whatsapp.
I expected it to say "Kafka" but given how far back it was, ejabberd makes perfect sense! It must have been amazing to work at such a high-growth, engineering-focused company building such a large public-facing service, at that time in internet history
Feature creep can be harder than people think. One of the reasons why Linux is a thing, was Linus understanding that there has to be someone with power to say "No" to anything proposed.
i love the impending feeling of i'm doom i get from watching your videos. because i have no clue what 90% of the tings are your speaking about and will looking for a job in like a year
I love that he got rejected by Facebook, then they acquired his company --7 years later for billions... I've heard it described before as the "most expensive hire ever" 😅
another thing about yahoo engeneers , of all the big name .commers, they were the only ones toe also register bitflip domain names , where the rest only rigistered the typo domain names ... they foresaw and closed that vulnerability long before anyone ever came up with abusing it
Turns out (as somebody mentioned in the chat) that Jan Koum is actually Ukrainian. It's the English transliteration that makes it look absolutely non-Ukrainian (I thought he's Chinese, lol) P.S. You read the name correctly, and the surname -is more like Koom ("oo" as in "door")-
1:40 I still don't understand why YUI was not maintained. I mean NodeJS and React and all that are good, but nothing beats a well designed, single-source comprehensive JS framework. It was the JDK of Javascript.
Simple: make a stable application on stable server platforms. You really don’t need thousands of engineers for applications that are that simple. I struggle to understand why even Netflix requires hundreds of engineers full time.
everyone has green threads because they were copying Erlang, even the Go guys knew about Erlang but Erlang is even better than Go, Rob Pike also recommends Erlang. Erlang code always lives in a green thread, and all of their green threads are individual erlang runtimes with their own heap so when you kill one of them you are automatically collecting garbage, giving Erlang "arenas" by default. Erlang's hot reloading works due to always running on a green thread, because they can load v1 and v1.1 in different threads, last time I checked it had only support for 2 simultaneous versions.
Maximum of 10,000 Clients times 60 times per second = 600,000 UDP messages per second with my game server. Admittedly, it's just a multiplayer dungeon crawler, so very small packets, and no wasting time with TCP. I just want to know why it took them 32 engineers ;)
It looks like AI generated (no offense). Looked like those generic "best practises" articles... You could've made it more technical provided you had more information at hand.
Yahoo were pioneers in scaling early day web setvices, so the statement about how early day engineers were among the greatest of their time doesn't surprise me
When you and your family are LatAm, you cant really say no to Whatsapp unfortunately. Signal is awesome, but god be damned if you aren't in the family whatsapp for your grandma's daily drop of prayer chains
SMS were so expensive in Brazil back then that even now that it is cheap people only use WhatsApp. And some of less tech ones savvy don’t event know what an sms is
Whatsapp should be using the Actor Model with high concurrency, similar to Akka, I participated in an HR interview for a Reactive position and they told me they use Erlang in most of their stack services. The Actor model is so interesting as a design pattern an I think should be spread more among us mortals.
We have a very serious communication problem in our "team" right now. I put "team" in quotes because it's actually about 50 people in six work streams that are identical. It's so bad that I can't make any breaking changes to our API at all since it is practically IMPOSSIBLE to communicate with everyone involved in a certain feature since there is no central person who knows who is working on what since the six work streams actually all do the same thing and work on the same code base.
My dude....they may handle 50 billion messages.... but for over a year, the wrong avatar picture of the wrong person showed up in a convo using the desktop app.
